mt/am_matmul/matmul_mi.c~

   1 //**************************************************************************
   2 // Multi-threaded Matrix Multiply benchmark
   3 //--------------------------------------------------------------------------
   4 // TA     : Christopher Celio
   5 // Student:
   6 //
   7 //
   8 // This benchmark multiplies two 2-D arrays together and writes the results to
   9 // a third vector. The input data (and reference data) should be generated
  10 // using the matmul_gendata.pl perl script and dumped to a file named
  11 // dataset.h.
  12
  13
  14 // print out arrays, etc.
  15 //#define DEBUG
  16
  17 //--------------------------------------------------------------------------
  18 // Includes
  19
  20 #include <string.h>
  21 #include <stdlib.h>
  22 #include <stdio.h>
  23
  24
  25 //--------------------------------------------------------------------------
  26 // Input/Reference Data
  27
  28 typedef float data_t;
  29 #include "dataset.h"
  30
  31
  32 //--------------------------------------------------------------------------
  33 // Basic Utilities and Multi-thread Support
  34
  35 __thread unsigned long coreid;
  36 unsigned long ncores;
  37
  38 #include "util.h"
  39
  40 #define stringify_1(s) #s
  41 #define stringify(s) stringify_1(s)
  42 #define stats(code) do { \
  43     unsigned long _c = -rdcycle(), _i = -rdinstret(); \
  44     code; \
  45     _c += rdcycle(), _i += rdinstret(); \
  46     if (coreid == 0) \
  47       printf("%s: %ld cycles, %ld.%ld cycles/iter, %ld.%ld CPI\n", \
  48              stringify(code), _c, _c/DIM_SIZE/DIM_SIZE/DIM_SIZE, 10*_c/DIM_SIZE/DIM_SIZE/DIM_SIZE%10, _c/_i, 10*_c/_i%10); \
  49   } while(0)
  50
  51
  52 //--------------------------------------------------------------------------
  53 // Helper functions
  54
  55 void printArray( char name[], int n, data_t arr[] )
  56 {
  57    int i;
  58    if (coreid != 0)
  59       return;
  60
  61    printf( " %10s :", name );
  62    for ( i = 0; i < n; i++ )
  63       printf( " %3ld ", (long) arr[i] );
  64    printf( "\n" );
  65 }
  66
  67 void __attribute__((noinline)) verify(size_t n, const data_t* test, const data_t* correct)
  68 {
  69    if (coreid != 0)
  70       return;
  71
  72    size_t i;
  73    for (i = 0; i < n; i++)
  74    {
  75       if (test[i] != correct[i])
  76       {
  77          printf("FAILED test[%d]= %3ld, correct[%d]= %3ld\n",
  78             i, (long)test[i], i, (long)correct[i]);
  79          exit(-1);
  80       }
  81    }
  82
  83    return;
  84 }
  85
  86 //--------------------------------------------------------------------------
  87 // matmul function
  88
  89 // single-thread, naive version
  90 void __attribute__((noinline)) matmul_naive(const int lda,  const data_t A[], const data_t B[], data_t C[] )
  91 {
  92    int i, j, k;
  93
  94    if (coreid > 0)
  95       return;
  96
  97    for ( i = 0; i < lda; i++ )
  98       for ( j = 0; j < lda; j++ )
  99       {
 100          for ( k = 0; k < lda; k++ )
 101          {
 102             C[i + j*lda] += A[j*lda + k] * B[k*lda + i];
 103          }
 104       }
 105
 106 }
 107
 108
 109
 110 void __attribute__((noinline)) matmul(const int lda,  const data_t A[], const data_t B[], data_t C[] )
 111 {
 112   /*size_t i, j, k;
 113   int row, column, column2, column3, column4, column5, column6, column7, column8;
 114   size_t max_dim = 32*32;
 115   data_t element, element2, element3, element4, element5, element6, element7, element8;
 116   data_t temp_mat[32]={0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
 117   for (i=coreid*max_dim/ncores; i<(max_dim/ncores+coreid*max_dim/ncores); i+=8){
 118     element=A[i];
 119     element2 = A[i+1];
 120     element3 = A[i+2];
 121     element4 = A[i+3];
 122     element5 = A[i+4];
 123     element6 = A[i+5];
 124     element7 = A[i+6];
 125     element8 = A[i+7];
 126     row= (int)(i/32)*32;
 127     column = i%32*32;
 128     column2 = (i+1)%32*32;
 129     column3 = (i+2)%32*32;
 130     column4 = (i+3)%32*32;
 131     column5 = (i+4)%32*32;
 132     column6 = (i+5)%32*32;
 133     column7 = (i+6)%32*32;
 134     column8 = (i+7)%32*32;
 135
 136     for (j=0; j<32; j+=8){
 137       temp_mat[j]+=element*B[column+j]+element2*B[column2+j]+element3*B[column3+j]+element4*B[column4+j]+element5*B[column5+j]+element6*B[column6+j]+element7*B[column7+j]+element8*B[column8+j];
 138       temp_mat[j+1]+=element*B[column+j+1]+element2*B[column2+j+1]+element3*B[column3+j+1]+element4*B[column4+j+1]+element5*B[column5+j+1]+element6*B[column6+j+1]+element7*B[column7+j+1]+element8*B[column8+j+1];
 139       temp_mat[j+2]+=element*B[column+j+2]+element2*B[column2+j+2]+element3*B[column3+j+2]+element4*B[column4+j+2]+element5*B[column5+j+2]+element6*B[column6+j+2]+element7*B[column7+j+2]+element8*B[column8+j+2];
 140       temp_mat[j+3]+=element*B[column+j+3]+element2*B[column2+j+3]+element3*B[column3+j+3]+element4*B[column4+j+3]+element5*B[column5+j+3]+element6*B[column6+j+3]+element7*B[column7+j+3]+element8*B[column8+j+3];
 141       temp_mat[j+4]+=element*B[column+j+4]+element2*B[column2+j+4]+element3*B[column3+j+4]+element4*B[column4+j+4]+element5*B[column5+j+4]+element6*B[column6+j+4]+element7*B[column7+j+4]+element8*B[column8+j+4];
 142       temp_mat[j+5]+=element*B[column+j+5]+element2*B[column2+j+5]+element3*B[column3+j+5]+element4*B[column4+j+5]+element5*B[column5+j+5]+element6*B[column6+j+5]+element7*B[column7+j+5]+element8*B[column8+j+5];
 143       temp_mat[j+6]+=element*B[column+j+6]+element2*B[column2+j+6]+element3*B[column3+j+6]+element4*B[column4+j+6]+element5*B[column5+j+6]+element6*B[column6+j+6]+element7*B[column7+j+6]+element8*B[column8+j+6];
 144       temp_mat[j+7]+=element*B[column+j+7]+element2*B[column2+j+7]+element3*B[column3+j+7]+element4*B[column4+j+7]+element5*B[column5+j+7]+element6*B[column6+j+7]+element7*B[column7+j+7]+element8*B[column8+j+7];
 145     }
 146     if (i%32==24){
 147       for(k=0; k<32; k++){
 148         C[row+k]=temp_mat[k];
 149         temp_mat[k]=0;
 150       }
 151     }
 152     }*/
 153     int i,j,k,l;
 154     data_t element1, element2, element3, element4, element5, element6, element7, element8;
 155     int row, row2;
 156     int column1, column2, column3, column4, column5, column6, column7, column8;
 157     data_t temp[32]={0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
 158     data_t temp2[32]={0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
 159     if (coreid == 0){
 160       for (i=0; i<32; i+=2){
 161         row = i*32;
 162         row2 = (i+1)*32;
 163         for (j=0; j<16; j+=4){
 164           element1 = A[row+j];
 165           element2 = A[row+j+1];
 166           element3 = A[row+j+2];
 167           element4 = A[row+j+3];
 168           column1 = j*32;
 169           column2 = (j+1)*32;
 170           column3 = (j+2)*32;
 171           column4 = (j+3)*32;
 172           element5 = A[row2+j];
 173           element6 = A[row2+j+1];
 174           element7 = A[row2+j+2];
 175           element8 = A[row2+j+3];
 176
 177           for (k=0; k<32; k+=4){
 178             temp[k]+=element1*B[column1+k]+element2*B[column2+k]+element3*B[column3+k]+element4*B[column4+k];
 179             temp[k+1]+=element1*B[column1+k+1]+element2*B[column2+k+1]+element3*B[column3+k+1]+element4*B[column4+k+1];
 180             temp[k+2]+=element1*B[column1+k+2]+element2*B[column2+k+2]+element3*B[column3+k+2]+element4*B[column4+k+2];
 181             temp[k+3]+=element1*B[column1+k+3]+element2*B[column2+k+3]+element3*B[column3+k+3]+element4*B[column4+k+3];
 182             temp2[k]+=element5*B[column1+k]+element6*B[column2+k]+element7*B[column3+k]+element8*B[column4+k];
 183             temp2[k+1]+=element5*B[column1+k+1]+element6*B[column2+k+1]+element7*B[column3+k+1]+element8*B[column4+k+1];
 184             temp2[k+2]+=element5*B[column1+k+2]+element6*B[column2+k+2]+element7*B[column3+k+2]+element8*B[column4+k+2];
 185             temp2[k+3]+=element5*B[column1+k+3]+element6*B[column2+k+3]+element7*B[column3+k+3]+element8*B[column4+k+3];
 186           }
 187           if (j==12){
 188             for (l=0; l<32; l++){
 189               C[row+l]+=temp[l];
 190               C[row2+l]+=temp2[l];
 191               temp[l]=0;
 192               temp2[l]=0;
 193             }
 194           }
 195         }
 196       }
 197     }
 198     else if (coreid==1){
 199       for (i=0; i<32; i+=2){
 200         row = (31-i)*32;
 201         row2 = (31-i-1)*32;
 202         for (j=16; j<32; j+=4){
 203           element1 = A[row+j];
 204           element2 = A[row+j+1];
 205           element3 = A[row+j+2];
 206           element4 = A[row+j+3];
 207           element5 = A[row2+j];
 208           element6 = A[row2+j+1];
 209           element7 = A[row2+j+2];
 210           element8 = A[row2+j+3];
 211           column1 = j*32;
 212           column2 = (j+1)*32;
 213           column3 = (j+2)*32;
 214           column4 = (j+3)*32;
 215           for (k=0; k<32; k+=4){
 216             temp[k]+=element1*B[column1+k]+element2*B[column2+k]+element3*B[column3+k]+element4*B[column4+k];
 217             temp[k+1]+=element1*B[column1+k+1]+element2*B[column2+k+1]+element3*B[column3+k+1]+element4*B[column4+k+1];
 218             temp[k+2]+=element1*B[column1+k+2]+element2*B[column2+k+2]+element3*B[column3+k+2]+element4*B[column4+k+2];
 219             temp[k+3]+=element1*B[column1+k+3]+element2*B[column2+k+3]+element3*B[column3+k+3]+element4*B[column4+k+3];
 220             temp2[k]+=element5*B[column1+k]+element6*B[column2+k]+element7*B[column3+k]+element8*B[column4+k];
 221             temp2[k+1]+=element5*B[column1+k+1]+element6*B[column2+k+1]+element7*B[column3+k+1]+element8*B[column4+k+1];
 222             temp2[k+2]+=element5*B[column1+k+2]+element6*B[column2+k+2]+element7*B[column3+k+2]+element8*B[column4+k+2];
 223             temp2[k+3]+=element5*B[column1+k+3]+element6*B[column2+k+3]+element7*B[column3+k+3]+element8*B[column4+k+3];
 224           }
 225           if (j==28){
 226             for (l=0; l<32; l++){
 227               C[row+l]+=temp[l];
 228               C[row2+l]+=temp2[l];
 229               temp[l]=0;
 230               temp2[l]=0;
 231             }
 232           }
 233         }
 234       }
 235       }
 236    // ***************************** //
 237    // **** ADD YOUR CODE HERE ***** //
 238    // ***************************** //
 239    //
 240    // feel free to make a separate function for MI and MSI versions.
 241
 242 }
 243
 244 //--------------------------------------------------------------------------
 245 // Main
 246 //
 247 // all threads start executing thread_entry(). Use their "coreid" to
 248 // differentiate between threads (each thread is running on a separate core).
 249
 250 void thread_entry(int cid, int nc)
 251 {
 252    coreid = cid;
 253    ncores = nc;
 254
 255    // static allocates data in the binary, which is visible to both threads
 256    static data_t results_data[ARRAY_SIZE];
 257
 258
 259    // Execute the provided, naive matmul
 260    barrier();
 261    stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 262
 263
 264    // verify
 265    verify(ARRAY_SIZE, results_data, verify_data);
 266
 267    // clear results from the first trial
 268    size_t i;
 269    if (coreid == 0)
 270       for (i=0; i < ARRAY_SIZE; i++)
 271          results_data[i] = 0;
 272    barrier();
 273
 274
 275    // Execute your faster matmul
 276    barrier();
 277    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier());
 278
 279 #ifdef DEBUG
 280    printArray("results:", ARRAY_SIZE, results_data);
 281    printArray("verify :", ARRAY_SIZE, verify_data);
 282 #endif
 283
 284    // verify
 285    verify(ARRAY_SIZE, results_data, verify_data);
 286    barrier();
 287
 288    exit(0);
 289 }
 290